PDF Extract — User-Defined Evidence Tables

Overview

The free analog of Elicit's "columns" feature. Point it at a stack of PDFs, declare the columns you care about, and get back a tidy evidence table — one row per paper, one column per question — as CSV or a Markdown table. Built for data extraction across a corpus: literature screening, methods comparison, PICO/sample-size pulls, and review evidence sheets.

Every input is normalized to clean Markdown via Microsoft MarkItDown (see the alterlab-markitdown skill) before any column is extracted, so PDF, DOCX, PPTX, HTML, and the other MarkItDown formats all work.

When to Use

• Extracting the same fields from many papers into a comparison table or spreadsheet.
• Building a data-extraction sheet for a review (sample size, design, intervention,

outcome, effect size, limitations).

• Screening a folder of PDFs into a CSV you can sort/filter.
• Any "make me a table where each row is a paper and each column is X" request.

Not for: searching databases to find papers, or synthesizing a narrative review. Discovering literature and writing the prose synthesis is alterlab-literature-review's job — this skill operates on PDFs you already have and produces a structured table, not prose. Use it after you have the papers.

Quick Start

SCRIPT=skills/research-tools/alterlab-pdf-extract/scripts/extract_to_table.py

# Heuristic backend: use 'label:question' so the question carries the words the
# papers actually use — overlap is lexical, so a bare label like "Sample size"
# rarely matches "we enrolled 240 participants" (see Backends below).
uv run python "$SCRIPT" papers/*.pdf \
    --column "Sample size:how many participants were enrolled" \
    --column "Main finding:primary result reported" \
    --column "Limitations:study limitations weaknesses" --format md

# Columns with explicit extraction questions, CSV to a file:
uv run python "$SCRIPT" a.pdf b.pdf \
    --column "n:how many participants were enrolled" \
    --column "effect:what was the main reported effect size" \
    --format csv -o evidence.csv

# Precise LLM-backed extraction (needs OPENROUTER_API_KEY) — label-only columns
# are fine here, since the model reads the paper rather than matching keywords:
uv run python "$SCRIPT" lit/*.pdf \
    --columns "Population" "Intervention" "Outcome" --backend llm

Defining Columns

A column is a label plus the question that drives extraction. Two ways to declare:

Use --columns for a space-separated list of simple labels, and repeat --column for specs that need an explicit question. Both combine; duplicate labels are rejected.

Backends

• heuristic (default, offline, free): for each column, scores Markdown sentences by

literal keyword overlap with the column spec and emits the best evidence snippet. Deterministic — good for a first pass, screening, and reproducible runs. Returns an empty cell when no sentence shares a keyword. Gotcha: matching is lexical, not semantic. A generic label like "Sample size" tokenizes to sample, size — which never appears in "we enrolled 240 participants", so the cell comes back empty. For the heuristic backend, phrase columns as Label:question where the question uses the words the papers actually use (participants, enrolled, cohort, n=). Quantitative questions also get a small ranking bonus for sentences containing digits when the question itself contains a digit or %. When you only have generic labels, use the llm backend instead.

• llm (optional, precise): sends the converted Markdown plus all column questions to an

OpenAI-compatible endpoint (OpenRouter by default) for a concise per-column answer (N/A when the paper is silent). The model ID follows the ALTERLAB_MODEL convention (skills/core/shared/model_env.md): $ALTERLAB_MODEL, else the dated default. Override per-run with --model.

Output

• --format md → GitHub-flavored Markdown table (pipes in cell text are escaped).
• --format csv → CSV via csv.DictWriter.
• -o FILE writes to disk (UTF-8); otherwise the table goes to stdout.

The first column is always Source (the input filename); a file that fails conversion yields an ERROR row so the table stays aligned.

Pipeline Fit

alterlab-pdf-extract sits downstream of paper discovery and upstream of synthesis:

1. Find papers — database/search skills (alterlab-literature-review, search backends).
2. Extract the evidence table — this skill.
3. Synthesize / write up — alterlab-literature-review, alterlab-paper-writer.

Files

• scripts/extract_to_table.py — the CLI. --help lists every flag; --columns /

--column define the table; --backend, --format, --model, -o control extraction and output.